Research map / Label-efficient and robust learning / Semi-supervised learning
MATCH: Multi-faceted Adaptive Topo-Consistency for Semi-Supervised Histopathology Segmentation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MATCH, a semi-supervised histopathology segmentation framework that enforces topological consistency across MC-dropout predictions and temporal training snapshots, using a new spatial-overlap-plus-persistence Hungarian matching (MATCH-Pair/MATCH-Global).
Problem
Semi-supervised segmentation of densely packed histopathology structures (glands, nuclei) suffers topological errors such as false merges and splits, and existing topology-aware SSL methods rely on hand-picked persistence thresholds or matching that ignores spatial correspondence when ground truth i
Approach
Dual-level topological consistency: persistent-homology features (0-D) are matched across MC-dropout facets and across temporal snapshots with a Hungarian matching cost combining IoU overlap, normalized persistence weight, and spatial proximity; matched features are pushed toward confident birth/death probabilities while unmatched ones are shrunk toward zero lifespan.
Key results
- On CRAG with 20% labels, MATCH reaches Dice_obj 0.909 and BE 0.188 versus TopoSemiSeg's 0.898 and 0.226.
- On GlaS 10% labels, MATCH improves BME from 8.300 (TopoSemiSeg) to 7.850 and DIU from 35.845 to 30.525, while Dice is 0.884 vs 0.878.
- On MoNuSeg 20% labels, MATCH is slightly below TopoSemiSeg on Dice_obj (0.790 vs 0.793) but better on BE, BME and DIU (4.930, 179.225, 982.286 vs 5.150, 188.642, 1105.946).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MATCH instead of TopoSemiSeg | Semi-supervised segmentation · CRAG 20% labeled | BE: 0.226 $\pm$ 0.019 -> 0.188 $\pm$ 0.018 | Helped |
| Use MATCH instead of TopoSemiSeg | Semi-supervised segmentation · GlaS 20% labeled | Dice_obj: 0.895 $\pm$ 0.003 -> 0.894 $\pm$ 0.004 | No clear effect |
| Add spatial proximity to the matching cost (ablation) | MATCH matching algorithm · CRAG 20% labeled | Dice_obj: 0.890 $\pm$ 0.005 -> 0.909 $\pm$ 0.005 | Helped |
| Add IoU to the matching cost (ablation) | MATCH matching algorithm · CRAG 20% labeled | Dice_obj: 0.882 $\pm$ 0.006 -> 0.909 $\pm$ 0.005 | Helped |
| Set B_intra and B_temp to 5 instead of 4 (ablation) | MATCH · CRAG 20% labeled | Dice_obj: 0.909 $\pm$ 0.005 -> 0.872 $\pm$ 0.011 | Hurt |
| Set B_intra and B_temp to 2 instead of 4 (ablation) | MATCH · CRAG 20% labeled | Dice_obj: 0.909 $\pm$ 0.005 -> 0.878 $\pm$ 0.010 | Hurt |
| Add intra-topological consistency loss (ablation) | MATCH · CRAG 20% labeled | Dice_obj: 0.862 $\pm$ 0.011 -> 0.898 $\pm$ 0.006 | Helped |
| Add temporal-topological consistency loss (ablation) | MATCH · CRAG 20% labeled | Dice_obj: 0.862 $\pm$ 0.011 -> 0.882 $\pm$ 0.008 | Helped |
| Replace MC dropout with variational inference (ablation) | MATCH · CRAG 20% labeled | Dice_obj: 0.909 $\pm$ 0.005 -> 0.895 $\pm$ 0.006 | Hurt |
| Replace MC dropout with temperature scaling (ablation) | MATCH · CRAG 20% labeled | Dice_obj: 0.909 $\pm$ 0.005 -> 0.891 $\pm$ 0.007 | Hurt |
| Remove pixel-wise consistency loss (ablation) | MATCH · CRAG 20% labeled | Dice_obj: 0.875 $\pm$ 0.008 -> 0.909 $\pm$ 0.005 | Hurt |
| Apply dual-level consistency between teacher and student instead of student-only (ablation) | MATCH · CRAG 20% labeled | Dice_obj: 0.909 $\pm$ 0.005 -> 0.885 $\pm$ 0.007 | Hurt |
Surprising findings
- On GlaS with 20% labels, TopoSemiSeg has the best Dice_obj (0.895 vs MATCH 0.894), so the claim of consistently superior pixel-wise performance does not hold across all settings.
- Using the teacher's multiple predictions for consistency performed worse than applying it to the student only (Dice 0.885 vs 0.909 on CRAG 20%).
Limitations
- Relies on stable 0-D feature extraction from persistence diagrams, which can fail under extreme noise or minimal differences between predictions.
- Extra computational overhead from multiple MC-dropout and snapshot forward passes and matching steps, increasing training time and memory.
- Pixel-wise gains over the strongest prior method are small or absent in some settings.
Experimental setup
Datasets: CRAG GlaS MoNuSeg Roads MoNuSAC
Benchmarks: CRAG GlaS MoNuSeg Roads MoNuSAC
Metrics: Dice_obj Betti error Betti matching error DIU VOI PCC
Baselines: MT EM UA-MT URPC XNet PMT TopoSemiSeg Wasserstein matching Betti matching
Builds on: TopoSemiSeg Betti Matching Mean Teacher UNet
Closest papers by meaning
- Weakly Supervised Gland Segmentation with Class Semantic Consistency and Purified Labels Filtration (AAAI 2025 · similarity 0.86)
- Prototype-Based Image Prompting for Weakly Supervised Histopathological Image Segmentation (CVPR 2025 · similarity 0.86)
- Incremental Nuclei Segmentation from Histopathological Images via Future-class Awareness and Compatibility-inspired Distillation (CVPR 2024 · similarity 0.86)
- Semantic and Visual Crop-Guided Diffusion Models for Heterogeneous Tissue Synthesis in Histopathology (NeurIPS 2025 · similarity 0.85)
- RankMatch: Exploring the Better Consistency Regularization for Semi-supervised Semantic Segmentation (CVPR 2024 · similarity 0.85)
- Anatomical Knowledge Mining and Matching for Semi-supervised Medical Multi-structure Detection (AAAI 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.